* Fortify SED_PG_CONF() shell function, as called by gpinitsystem

It's also called via gpcreateseg.sh

There is a chance where gpinitsystem may miss a GUC change in one of the
segments and left undetected. This would cause failures down the line.

More robust in three ways:
- Retry ssh calls in case they fail
- Actually error exit if it does fail, for quicker diagnosis if the
  value in postgres.conf did not get updated
- When keeping a version of the key-value pair in a comment, only
  archive the first-matched key-value (otherwise you get comments on
  your comments)

Use trap for retrying failed ssh
Use trap in a limited scope for now